NLP researcher at Google spanning multimodal in-context learning, LLM evaluation validity, and efficient MoE translation
5
papers here
4
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-17) and cover the author's entire public record, not just the papers listed here. Reproducibility signal comes from Evolve's own reviews.
Research profile
Output—
Not yet scored: this axis measures publication volume — yearly paper counts, total outputs, and authorship positions — which requires a fuller bibliographic record than the three reviewed arXiv papers in this snapshot provide.
Impact—
Not yet scored: this axis measures citation counts, h-index, and downstream influence on the field, none of which are present in the current snapshot for either the reviewed papers or the author overall.
Focus0/100
Scores low (0) because the reviewed record spreads evenly across three distinct fields — multimodal vision-language models, core NLP evaluation, and machine translation with model compression — each accounting for exactly one third of the three reviewed papers. With only one paper per area, this reads as even topical distribution rather than confirmed fragmentation, since a single paper per field cannot establish sustained breadth or depth.
Collaboration68/100
Moderate (55): the snapshot records 15 distinct coauthors across the reviewed work, with a recurring cluster that includes Zihan Xue, Po-Yi Lu, Serhii Honcharenko, and Zih-Ching Chen. All three papers are multi-authored 2026 arXiv submissions with Peng in a middle or last position and no corresponding-author claim, indicating steady team-based collaboration; the mid-range score reflects a real but thinly documented network.
Activity—
Not yet scored: this axis measures publication cadence and year-over-year research output; the snapshot shows only three papers, all dated 2026, so a sustained activity rate cannot be estimated.
Reproducibility66/100
Moderate (55): Evolve assessed three papers, averaging 55 with each rated medium readiness. The WMT26 system paper is clear and honest but reports only internal pseudo-reference evaluation without ablations; TwinICL has strong controls yet toy shape tasks and partial run coverage; the judge-audit is well hedged but small, with 20 premises and 7 readers.
Research summary
Nanyun Peng is a researcher at Google whose reviewed work spans three adjacent areas of natural language processing and machine learning: multimodal vision-language modeling, evaluation methodology for large language models, and efficient machine translation systems. The Evolve catalog includes three 2026 arXiv papers, in each of which Peng appears as a middle or last author with no corresponding-author claim. TwinICL introduces a paired-counterfactual benchmark that exposes persistent multimodal in-context learning gaps across six vision-language models. An ESTS WMT26 system paper prunes a GPT-OSS-20B mixture-of-experts model into 4.2–7.8B translation specialists using routing mass. A third study audits a psychological-depth LLM judge that flips to 89% pro-reasoning on frontier story pairs where human readers split, despite a development correlation of ρ=0.646. Bibliometrically, the snapshot records 15 distinct coauthors and an even one-third field split across the three areas, producing a low focus score (0) and a moderate collaboration score (55). Reproducibility averages 55 over three assessed papers, all rated medium readiness, with recurring limitations: partial run coverage, small human-study samples of 20 premises and 7 readers, and missing ablations. Citation counts, h-index, and annual output are unavailable in this snapshot, so impact, output, and activity remain unscored rather than estimated.
Collaboration pattern
The reviewed work is consistently multi-authored: across three 2026 arXiv papers, Peng appears as a middle or last author with no corresponding-author claim. The snapshot records 15 distinct coauthors, with a recurring group including Zihan Xue, Po-Yi Lu, Serhii Honcharenko, Zih-Ching Chen, and Hsuan-Tien Lin, suggesting stable small-team collaboration within Google and with academic partners. The moderate collaboration score (55) reflects a real but narrowly documented network; co-author seniority and full affiliation data are unavailable.
Research keywords
Highlights
Paired-counterfactual benchmark exposes multimodal in-context learning gaps
LLM judge of psychological depth tilts toward reasoning models
Routing-informed expert pruning yields compact WMT26 translation systems
Timeline
- 2026Published TwinICL, a paired-counterfactual benchmark diagnosing multimodal in-context learning across six vision-language models.
- 2026Co-authored the ESTS WMT26 paper on routing-informed pruning of GPT-OSS-20B into 4.2–7.8B translation specialists.
- 2026Co-authored an audit showing an LLM psychological-depth judge flips 89% pro-reasoning where human readers split.
Papers in the Evolve catalog
TwinICL: Diagnosing Multimodal In-Context Learning through Paired Counterfactuals
Does Reasoning Improve Psychological Depth in Large Language Models? It Depends on Who's Judging
ESTS at WMT26: Routing-Informed Expert Pruning for Model Compression
NormViz: A Benchmark and Framework for Grounding Multimodal Reasoning in Global Cultures
Frequent co-authors
About this profile
This profile is AI-structured by Evolve.Pub from public bibliographic data (OpenAlex) and Evolve's own paper reviews. It is not written or endorsed by the researcher unless the profile is claimed. Radar dimensions are computed from real bibliometrics with fixed formulas — raw figures are shown alongside; missing dimensions mean insufficient data, never zero. Errors? Claim the profile to correct or hide it.
- Bibliometrics (coauthor count, field distribution, and radar scores) come from public bibliographic data only; no citation counts or h-index are available in this snapshot.
- Identity is name-resolved from public bibliographic data; biographical details beyond the reviewed papers and listed coauthors are not verified.
- All three reviewed papers are 2026 arXiv preprints and may not have completed peer review; reproducibility reflects Evolve's assessment of these three papers only.
- No ORCID or country-level affiliation data was available for identity confirmation.